baytorch 0.0.1.dev0

BayTorch - Bayesian layers for PyTorch

BayTorch - Bayesian layers in PyTorch

Bayesian layers are probabilistic layers where weights are stochastic and sampled from the posterior distribution or an approximation.

So far, this module includes diagonal Gaussian linear and convolutional layers with mean and standard deviation of every weight as learnable parameters, as well as an implementation of the variational free energy loss for stochastic gradient VI according to the Bayes-by-Backprop algorithm [1].

[1] Blundell et al., Weight Uncertainty in Neural Networks, 2015.

Installation

Install from PyPI

pip install baytorch

or clone the repository, build and install from source

git clone https://jugit.fz-juelich.de/ias-8/baytorch.git && cd baytorch
python -m build
pip install dist/baytorch*whl